Accelerating Discovery of Polyimides with Intrinsic Microporosity for Membrane‐Based Gas Separation: Synergizing Physics‐Informed Performance Metrics and Active Learning

Abstract Polymer membranes are widely used for gas separation, addressing critical environmental and energy issues. Nevertheless, crafting high‐performance polymer membranes remains a challenging task, often relies on labor‐intensive trial‐and‐error experiments. Different from the traditional Edisonian approach, machine learning offers significant potential to expedite the design of new polymer membranes for gas separation, yet it faces a substantial hurdle due to limited data availability. To overcome this challenge, in this study two physics‐informed performance metrics are introduced, namely fractional free volume and average void size, to assess the separation performance of polymer membranes. By employing active learning and multi‐target screening, top‐performing polyimides with intrinsic microporosity are efficiently discovered from a pool of 155 610 candidates, through calculations of only 709 (0.45%) in the entire search space. As validated by molecular simulations, the top‐performing polyimides exhibit exceptional separation performance for CO 2 /N 2 , CO 2 /CH 4 , and O 2 /N 2 mixtures, superior to existing polymers and surpassing the current upper bound. The utilization of physics‐informed performance metrics provides a novel strategy to advance the design and accelerate the discovery of new polymer membranes for gas separation and many other important applications.